DeepMind AI agents accelerate cryptanalysis: two weaknesses in legacy RSA-1024; 3.5× faster 1024-bit root vs prior run

Why it matters

【Key points】

• On 2026-11-06 (17:20Z) Google DeepMind / Pushmeet Kohli’s X Article “Advancing the frontier of cryptanalysis” reported human-guided AI agents stress-testing legacy 1024-bit RSA: (1) signature forgery—once an attacker controls ~half+ of the signed block, PKCS#1 v1.5-style padding adds essentially no protection
• shorter controlled regions get a ~1600× theoretical improvement
• (2) batch factoring—~30× lower average cost per additional same-length 1024-bit key.
• Two technical papers posted.
• Same Shea et al. 1024-bit instance: first-root CPU ~391 core-years vs ~1380 (~3.5×), run by two AI agents in ~4 days (peaks ~300k cores).
• Modern 2048-bit+ RSA unaffected; point is how fast AI rewrites hardness assumptions.

【Our read】

Background

hardness margins usually move over years; the story is agent research throughput (3.5× CPU, large engineering) plus PKCS#11 interfaces ranked by caller-controlled bits—math-copilot tempo applied to crypto assumptions.

Peers

OpenAI e/1089 is autonomous Linux LPE find/patch; DeepMind sits in number theory/standards and stresses 2048+ remains safe while hardness is no longer a slow variable for standards bodies.

Impact

no immediate mainstream threat; raises continuous stress-testing expectations and a citable half-block threshold for signing APIs. Watch official blog, affine-window demos at crypto size, and post-quantum agent work. paper

【Verifiable】

• Primary2026-11-06T17:20:31Z
• ArticlePDFs on storage.googleapis.com/deepmind-media/papers/ (Faster Affine Root, Fan Factories). No deepmind.google/blog card at patrol time. Not on site; ≠ OpenAI e/1089. max e/1092